KridgeDookie/KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 7, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

KridgeDookie/KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS is a 35.1 billion parameter Mixture of Experts (MoE) model with approximately 3 billion active parameters, derived from Kwaipilot/KAT-Coder-V2.5-Dev. This text-only model is specifically modified to significantly reduce refusal behavior while preserving its strong coding and agentic coding workflow capabilities. It offers a 32768 token context length and is optimized for developers seeking a highly compliant coding assistant.

Loading preview...

KAT-Coder V2.5 Dev 35B-A3B - ABLITERATED UNCENSORED PHILADELPHIA CLASS

This model is a specialized text-only Mixture of Experts (MoE) derivative, featuring 35 billion total parameters with approximately 3 billion active parameters. It is based on the Kwaipilot/KAT-Coder-V2.5-Dev and is specifically engineered to minimize refusal behavior through targeted post-training weight editing, while maintaining its core coding-oriented architecture and capabilities.

Key Capabilities & Differentiators

  • Significantly Reduced Refusals: Achieved 0 hard refusals across 842 internal test prompts and a separate 126-prompt holdout, demonstrating high compliance.
  • Coding & Agentic Workflows: Retains the strong coding capabilities of its parent model, making it suitable for various programming tasks.
  • High Coherence: Passed 23 out of 24 coherence checks in internal evaluations.
  • Flexible Deployment: Available in BF16 precision and various GGUF quantizations (Q4_K_M, Q5_K_M, Q8_0) for diverse hardware setups.
  • Extended Context: Supports a context length of 32768 tokens, beneficial for complex coding projects.

Use Cases

This model is ideal for developers and applications requiring a coding assistant that is highly resistant to refusals and can handle agentic coding workflows. It's particularly useful in environments where consistent, direct responses are preferred over cautious or evasive answers, especially for code generation and problem-solving. Users should review generated code for security and accuracy, as refusal reduction does not guarantee correctness or safety.

Finetunes
1 models